AI Input Contribution Ranking via Integrated Gradient Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence models struggle to explain how a group of inputs collaboratively contribute to an output, as current methods like SHAP and integrated gradients are limited in identifying collaborative input groups and are not scalable for interconnected models.
Innovation Solution
A resource conservation system that identifies a subset of inputs contributing most to an output by using a characterization output from an AI model, processing data elements through multiple models, and applying integrated gradient equations to determine the importance of each data element, allowing for the ranking and selection of the most impactful data elements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If integrated gradient equations are applied to identify important inputs in interconnected AI models, then measurement precision of input contribution is improved, but device complexity increases
Solution Approach 1:
The patent segments the explanation process into modular components by identifying individual important inputs within each interconnected AI model separately, then combining these segments to form a comprehensive explanation. This allows precise measurement of input contributions while managing complexity through systematic decomposition of the overall system.
2Reliability
If human intervention is used to recreate the input process, then reliability of explanation is improved, but loss of time increases
Solution Approach 1:
The patent implements self-service by enabling the AI system to automatically identify and explain important inputs using integrated gradient equations, eliminating the need for human intervention in the explanation process. The system performs self-explanation through computational methods that reliably identify input contributions without requiring manual recreation of the input process.
3Productivity
If all inputs are processed through multiple interconnected models, then productivity of analysis is improved, but use of energy increases
Solution Approach 1:
The patent extracts only the necessary information by identifying and isolating the most important inputs that contribute to the output, rather than processing all inputs through multiple interconnected models. This extraction approach maintains high productivity in analyzing critical factors while significantly reducing energy consumption by focusing computational resources only on essential inputs.
Data Source
AI summary
A resource conservation system, including a determination processor may be provided. The determination processor may identify a characterization output that characterizes a plurality of data structures. The characterization output may be based on plurality of inputs. The inputs may be processed through a plurality, or cascade, of artificial intelligence models both in sequence and in parallel. A numerical value may be identified for each data structure. The value may identify a degree of certainty that the determination processor accurately characterized each data structure. When the degree is above a threshold, the determination processor may identify a subset of inputs that most contributed to the characterization output. The determination processor may execute an equation to identify a subset of inputs that most contributed to the output. The equation may involve inputs and/or outputs of each of the cascade of models. Identified inputs may be ranked based on contribution to the outcome.